JP2015087556A - Voice recognition wfst generation device, voice recognition device, voice recognition wfst generation method, voice recognition method, program, and recording medium - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a voice recognition WFST generation device capable of suppressing a size of a voice recognition WFST.SOLUTION: A common unigram creation unit 11 reads frequencies of appearance of words from each of a plurality of language models 10to 10to calculate a common unigram value, and creates a common unigram WFST from the common unigram value. A first-stage WFST creation unit 16 performs a combining operation of M (M≥1) acoustic models WFST 13to WFST 13, a triphone WFST 14, a dictionary WFST 15, and the common unigram WFST, and connects each WFST for all acoustic models to constitute a first-stage WFST. A second-stage WFST creation unit 17 calculates an n-gram probability obtained by removing the common unigram value as an input from an n-gram probability of each language model while referring to the plural language models 10to 10to create a WFST of the n-gram probability, and connects each WFST of the n-gram probabilities of the respective language modes of all the language models 10to 10to constitute a second-stage WFST.

Description

  The present invention relates to a speech recognition WFST creation device for creating a weighted finite state transducer (hereinafter referred to as WFST) using a plurality of types of acoustic models and language models, a speech recognition device, a method, a program thereof, and a recording It relates to the medium.

  Speech recognition using WFST is converted into WFST that integrates information necessary for speech recognition, such as acoustic models, dictionaries, and language models, and the speech recognition target speech that is input is decoded with WFST regarded as a search space. This is a method of converting into a speech recognition result character string. Speech recognition using WFST is disclosed in Patent Documents 1 and 2, for example.

Japanese Patent No. 5175325 Japanese Patent No. 4478088

  In order to realize speech recognition using WFST with high recognition accuracy in a speech recognition service for various speakers, environments, and topics, it is necessary to simultaneously use a plurality of acoustic models and a plurality of language models. When the number of acoustic models and language models used simultaneously increases, the WFST size increases, and there is a problem that the memory required for speech recognition processing becomes enormous.

  The present invention has been made in view of such a problem, and a speech recognition WFST creation device capable of creating a WFST with reduced increase in memory size even when supporting a plurality of speech recognition services, a speech recognition device, and An object of the present invention is to provide a method, a program, and a recording medium.

  The speech recognition WFST creation apparatus of the present invention includes a common unigram WFST creation section, a common unigram WFST storage section, a first stage WFST creation section, and a second stage WFST creation section. The common unigram WFST creation unit reads the appearance frequency of each word from a plurality of language models, calculates a common unigram value, creates a common unigram WFST from the common unigram value, and creates the common unigram value and the common unigram WFST. Output. The common unigram WFST storage unit stores a common unigram value and a common unigram WFST. The first-stage WFST creation unit creates WFSTs for each acoustic model obtained by synthesizing N (N ≧ 1) acoustic model WFSTs, triphones WFST, dictionary WFST, and common unigram WFST. The first-stage WFSTs are configured by combining the WFSTs for each. The second-stage WFST creation unit receives the common unigram value, calculates each of the plurality of language models, calculates the n-gram probability obtained by removing the common unigram value with respect to the n-gram probability for each language model, and An n-gram probability WFST is created, and the n-gram probability WFSTs of all language models are combined to form a second-stage WFST.

  Further, the speech recognition apparatus of the present invention includes a first-stage WFST storage unit that stores the first-stage WFST created by the speech recognition WFST creation apparatus, a second-stage WFST storage unit that stores the second-stage WFST, A voice recognition unit. The speech recognition unit performs speech recognition by multi-stage on-the-fly synthesis with reference to the first stage WFST storage unit and the second stage WFST storage unit.

  According to the speech recognition WFST creation apparatus of the present invention, a common unigram WFST is created by reading the appearance frequency of each word from a plurality of language models and calculating a common unigram value. And since the 1st stage WFST shared by the common unigram for every acoustic model of a several acoustic model is comprised, the memory size of 1st stage WFST can be suppressed. In addition, since the speech recognition apparatus of the present invention performs speech recognition using the first stage WFST and the second stage WFST created by the speech recognition WFST creation apparatus described above, even when supporting a plurality of speech recognition services, Enables highly accurate speech recognition with a small amount of memory.

The figure which shows the function structural example of the WFST production apparatus 100 for speech recognition of this invention. The figure which shows the operation | movement flow of the WFST production apparatus 100 for speech recognition. The figure which shows the common unigram WFST of the word w. The figure which shows the example of 1st stage WFST. The figure which shows the function structural example of the speech recognition apparatus 200 of this invention.

  Embodiments of the present invention will be described below with reference to the drawings. The same reference numerals are given to the same components in a plurality of drawings, and the description will not be repeated.

  FIG. 1 shows a functional configuration example of a speech recognition WFST creation apparatus 100 of the present invention. The operation flow is shown in FIG. The speech recognition WFST creation apparatus 100 includes a common unigram WFST creation unit 11, a common unigram WFST storage unit 12, a first-stage WFST creation unit 16, and a second-stage WFST creation unit 17. The speech recognition WFST creation apparatus 100 is realized by a predetermined program being read into a computer composed of, for example, a ROM, a RAM, and a CPU, and the CPU executing the program.

The common unigram WFST creation unit 11 reads the appearance frequency of each word from a plurality of language models 10 ₁ , 10 ₂ ,..., 10 _N , calculates a common unigram value, and creates a common unigram WFST from the common unigram value. The common unigram value and the common unigram WFST are output. The language model 10 _1, for example the field of sports language model, the language model 10 _2, for example of the performing arts field language model, the language model 10 _N is a language model, for example, the political field. In this way, language models are prepared for a plurality of fields. Note that each language model includes appearance frequency information of words (only unigrams may be used).

The common unigram value (appearance probability) Pc (w) of the word w in the language model 10 _* (1 ≦ * ≦ N) is calculated by the following equation.

Here, C _* (w) is the appearance frequency of the word w in the language model 10 _* , and W is the number of words included in all the language models. The common unigram WFST creation unit 11 calculates and obtains a common unigram value (formula (1)) for all the words w to form a common unigram model. The common unigram model is a set (for example, a file) of common unigram values including all the words w.

  FIG. 3 shows a common unigram WFST. The common unigram WFST is composed of three nodes, and a state number is written in each node. The initial state is state 0. A state 2 surrounded by a double line represents an end state. An arc represents a state transition. <S> written together with an arc transitioning from state 0 to state 1 represents the beginning of a sentence. An arc that returns from state 1 and returns to state 1 indicates that word w is output with the appearance probability represented by a common unigram value in response to the input of word w. This arc is created for the number of words. </ S> written in the arc from state 1 to state 2 represents the end of the sentence. The common unigram WFST is stored in the common unigram WFST storage unit 12.

The first-stage WFST creation unit 16 includes M (M ≧ 1) acoustic models 13 ₁ , 13 ₂ ,... 13 _M , a triphone WFST 14, a dictionary WFST 15, and a common unigram WFST created by the common unigram creation unit 11. The first WFST is created by combining all the WFSTs for the respective acoustic models (step S16).

The WFST for each acoustic model 13 _* (1 ≦ * ≦ M) obtained by the synthesis operation is shown in the following equation.

Here, opt represents a WFST optimization operation, and ◯ represents a WFST composition operation. H _* is the WFST of each acoustic model. C is a triphone WFST, which converts the triphone output from the acoustic model WFST into phonemes. L is a dictionary WFST that converts phonemes into words. Gc is a common unigram WFST. A method for synthesizing and optimizing WFST is described, for example, in Reference 1 (Takaaki Hori, Mototsuka Tsukada, “Speech Recognition Using Weighted Finite State Transducers” Information Processing, October 15, 2004, Vol. 45, No. 10) As is well known.

The first-stage WFST creation unit 16 combines all WFST (H _* CLGc) for each synthesized and optimized acoustic model to form the first-stage WFST. FIG. 4 shows an example of the first stage WFST. The operation of the first stage WFST creation unit 13 will be described with reference to FIG.

The first stage WFST creation unit 13 creates a state s1 (initial state) and a state s2. Next, a transition is created in which both the input symbol and the output symbol are ε (empty) and are connected from the state s1 to the initial state of H ₁ CLGc including the acoustic model 1. Similarly, transitions from the state s1 are created for H ₂ CLGc,..., H _M CLGc corresponding to all acoustic models. Next, the input symbol and the output symbol are both ε, and a transition connecting from each end state of all H ₁ CLGc to H _M CLGc to the state s2 is created to constitute one first stage WFST.

In this way, the first stage WFST is configured by connecting WFST (H _* CLGc) for each acoustic model in parallel. The first stage WFST is output from the first stage WFST creating unit 16 to the outside. Alternatively, it may be stored in the first stage WFST storage unit 19.

The second stage WFST creation unit 17, the common inputs the common unigram value Pc (w), a plurality of language models 10 _1, 10 _2, with reference to ... 10 _N with respect to n-gram probability of each language model unigram An n-gram probability from which the value Pc (w) is removed is calculated to create a WFST having the n-gram probability, and a second-stage WFST in which the n-gram probabilities WFST of all language models are combined in parallel is created (step) S17). The n-gram probability Pc _* (w | uv) (u, v, w is a word) for each language model from which the common unigram value Pc (w) is removed can be calculated by the following equation.

The second-stage WFST creation unit 17 creates a WFST for each language model from the n-gram probabilities P _* (w | uv) for each language model from which the common unigram value Pc (w) is removed for all language models. . Then, the WFST for each language model is coupled in parallel to form a second stage WFST. The second stage WFST is a WFST based on the n-gram probability obtained by removing the common unigram value Pc (w) from the trigram probability P _* (w | uv) as is clear from the equation (3).

  The method of constructing the second stage WFST by combining the WFST for each language model in parallel is the same as the first stage WFST described with reference to FIG. The created second stage WFST is output to the outside. Alternatively, it may be stored in the second stage WFST storage unit 20.

  The processes of the common unigram creating unit 11, the first stage WFST creating unit 16, and the second stage WFST creating unit 17 described above are repeated until the processing for all acoustic models and language models is completed (No in step S18). . The control unit 18 performs the control of the time series operation and the control of the operation end in steps S11, S16, and S17. The function of the control unit 18 is not a special technical feature of this embodiment but a general one.

  The speech recognition WFST creation apparatus 100 described above can greatly reduce the size of the first stage WFST by using the common unigram WFST (Gc). Specifically, the number of first stage WFSTs can be reduced by N-1.

Although the common unigram creation unit 11 has been described as calculating the common unigram value by counting the appearance frequency of the word w for each language model, the common unigram value is obtained without using the appearance frequency of the word w. Anyway. In that case, the common unigram value Pc (w) is calculated by the following equation using the unigram probability P _* (w) of the word w originally possessed by the language model.

The method of creating the first stage WFST after obtaining the common unigram value Pc (w) is the same as described above. Since the unigram probability P _* (w) is a known value in each language model, it is not necessary to prepare the appearance frequency of the word w separately.

The above explanation has been made on the assumption that the words included in all language models 10 _* are the same. The words included in each language model 10 _* may be different. If the words included in the language model 10 _* are different, all the words included in the language model 10 _* are registered in the dictionary WFST15, and the calculation is performed using the formula (1) or the formula (4) for all the words. The common unigram value Pc (w) may be obtained. The method of creating the subsequent first stage WFST is the same as the method described above.

[Voice recognition device]
FIG. 5 shows a functional configuration example of the speech recognition apparatus 200 of the present invention. The speech recognition apparatus 200 includes a first stage WFST storage unit 19 that stores the first stage WFST created by the speech recognition WFST creation apparatus 100 described above, a second stage WFST storage unit 20 that stores the second stage WFST, And a voice recognition unit 210.

  The speech recognition unit 210 performs speech recognition by multi-stage on-the-fly synthesis using the first stage WFST and the second stage WFST stored in the first stage WFST storage unit 19 and the second stage WFST storage unit 20. . Since the speech recognition unit 210 searches for the first-stage WFST reduced in size by using the common unigram WFST (Gc) and the second-stage WFST that converts the common unigram into a trigram, the speech recognition processing is performed. Accurate speech recognition processing can be performed with a small memory size. Multistage on-the-fly speech recognition is described in Reference 2 (Takaaki Hori, Atsushi Nakamura “Generalized Fast On-the-fly Composition Algorithm fot WFST-Based Speech Recognition”, Proc. Of INTERSPEECH 2005.) Is well known.

  The speech recognition WFST creation apparatus 100 according to the present invention uses the WFST corresponding to each of the plurality of acoustic models as the first-stage WFST, and the first-stage WFST created for each acoustic model. The size of can be suppressed. Further, since the speech recognition apparatus 200 of the present invention performs multi-stage on-the-fly speech recognition using the first stage WFST and the second stage WFST created by the speech recognition WFST creation apparatus 100 of the present invention, a small memory It is possible to perform highly accurate speech recognition even with a size.

  When the processing means in the above apparatus is realized by a computer, the processing contents of the functions that each apparatus should have are described by a program. Then, by executing this program on the computer, the processing means in each apparatus is realized on the computer.

  The program is distributed by selling, transferring, or lending a portable recording medium such as a DVD or CD-ROM in which the program is recorded. Further, the program may be distributed by storing the program in a recording device of a server computer and transferring the program from the server computer to another computer via a network.

  Each means may be configured by executing a predetermined program on a computer, or at least a part of these processing contents may be realized by hardware.

Claims (8)

A common unigram WFST creation unit that reads the appearance frequency of each word from a plurality of language models, calculates a common unigram value, creates a common unigram WFST from the common unigram value, and outputs the common unigram value and the common unigram WFST When,
A common unigram WFST storage unit for storing the common unigram value and the common unigram WFST;
Create a WFST for each acoustic model obtained by synthesizing N (N ≧ 1) acoustic models WFST, triphone WFST, dictionary WFST and the common unigram WFST, and combine all WFSTs for each acoustic model. A first stage WFST creating unit constituting the first stage WFST;
Using the common unigram value as an input, referring to each of the plurality of language models, calculating the n-gram probability obtained by removing the common unigram value for the n-gram probability for each language model, and calculating the WFST of the n-gram probability. A second-stage WFST creating unit configured to create a second-stage WFST by combining the WFSTs of the n-gram probabilities of each language model of all language models,
A WFST creation apparatus for speech recognition comprising: In the WFST creation apparatus for speech recognition according to claim 1,
The common unigram WFST creation part
A speech recognition WFST creation apparatus, characterized in that the common unigram value is calculated without using the appearance frequency of words of the plurality of language models. A first-stage WFST storage section that stores the first-stage WFST created by the speech recognition WFST creation apparatus according to claim 1, and a second-stage WFST storage section that stores the second-stage WFST;
A speech recognition unit that performs speech recognition by multi-stage on-the-fly synthesis with reference to the first-stage WFST storage unit and the second-stage WFST storage unit;
A speech recognition apparatus comprising: A common unigram WFST creation process that reads the appearance frequency of each word from a plurality of language models, calculates a common unigram value, creates a common unigram WFST from the common unigram value, and outputs the common unigram value and the common unigram WFST When,
A WFST is created for each acoustic model obtained by synthesizing N (N ≧ 1) acoustic models WFST, triphone WFST, dictionary WFST, and the common unigram WFST, and all WFSTs for each acoustic model are combined. A first stage WFST creation process constituting the first stage WFST;
Using the common unigram value as an input, referring to each of the plurality of language models, calculating the n-gram probability obtained by removing the common unigram value for the n-gram probability for each language model, and calculating the WFST of the n-gram probability. A second-stage WFST creation process for creating a second-stage WFST by combining the WFSTs of the n-gram probabilities of each language model of all the language models,
A method for creating a speech recognition WFST. In the WFST creation method for voice recognition according to claim 4,
The common unigram WFST creation process is as follows:
A method for creating a speech recognition WFST, wherein the common unigram value is calculated without using the appearance frequency of words of the plurality of language models. A speech recognition process for performing speech recognition by multi-stage on-the-fly synthesis using the first stage WFST and the second stage WFST created by the speech recognition WFST creation method according to claim 4 or 5,
Including speech recognition method. A program for causing a computer to execute the function of each unit of the speech recognition WFST creation device according to claim 1 or 2, and the speech recognition device according to claim 3.   A computer-readable recording medium on which any one of the programs described in claim 7 is recorded.

Images